v0.5.0

[0.5.0] - 2026-06-02

Added

• Automatic memory hooks (all clients) — palace install now registers
  user-scope session start, post tool use, and stop hooks for every client
  that supports them, enforcing memory use in every project without per-project
  rules. post tool use auto-recalls relevant memory while the agent
  investigates; stop nudges the agent to record its work when it engaged Palace
  but saved nothing. Coverage:

  • Cursor — ~/.cursor/hooks.json (recall matched to Grep/Read).
  • Claude Code — ~/.claude/settings.json nested hooks (recall matched to
    Grep/Read/Glob).
  • Codex — ~/.codex/hooks.json nested hooks (recall matched to Bash,
    since Codex investigates through the shell; run /hooks once to trust them).
  • Claude Desktop — no hook system, so it remains rules-only (CLAUDE.md).

  Claude Code and Codex share a "Claude-style" output dialect
  (hookSpecificOutput.additionalContext; decision: "block" + reason on
  stop), selected via palace hook <event> --client <cursor|claude|codex>. The
  palace::hooks module exposes the pure, client-aware response builders.

• Cross-agent recall — palace_diary_search accepts all_agents and
  project_path to surface investigations recorded by any agent, and
  palace_session_context falls back to another agent's recent project work when
  the caller has none, so a different agent the next day still benefits from
  prior work.

Changed

• Write-path dedup — palace_diary_write (scoped per agent) and
  palace_remember now skip near-duplicate content, preventing diary/fact bloat
  from repeated similar writes. Durable decisions still belong in the
  deduplicated knowledge graph.

Fixed

• Resilient model download — the embedding model fetch now retries with
  exponential backoff, so transient HuggingFace failures (notably HTTP 429 rate
  limiting) no longer fail the first embedding call or CI. CI also caches the
  downloaded model (~/.cache/huggingface) across runs to avoid repeated
  downloads.